When many single-file uploads fire at once (ClaudeSync dispatch), they all
request the same project connection. One thread connects while the rest wait.
If that first attempt failed (usually transient rate-limiting), the waiters
raised a non-retryable 'another attempt already tried' error and popped a
modal — even though retrying would succeed.

Now the waiters raise a retryable ConnectionLostError, so with_retry backs
off (with jitter) and the thread attempts the connection itself. The jitter
desyncs the waiters so they don't all re-handshake simultaneously.